Learning Cross-Task Dependencies for Joint Extraction of Entities, Events, Event Arguments, and Relations

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9349 - 9360 December 7-11, 2022 ©2022 Association for Computational Linguistics Learning Cross-Task Dependencies for Joint Extraction of Entities, Events, Event Arguments, and Relations Minh Van Nguyen1, Bonan Min2∗, Franck Dernoncourt3, and Thien Huu Nguyen1,4 1 Dept. of Computer Science, University of Oregon, Eugene, OR, USA 2 Amazon AWS AI Labs 3 Adobe Research, Seattle, WA, USA 4 VinAI Research, Vietnam {minhnv,thien}@cs.uoregon.edu, bonanmin@amazon.com, dernonco@adobe.com Abstract Extracting entities, events, event arguments, and relations (i.e., task instances) from text rep- resents four main challenging tasks in informa- tion extraction (IE), which have been solved jointly (JointIE) to boost the overall perfor- mance for IE. As such, previous work often leverages two types of dependencies between the tasks, i.e., cross-instance and cross-type de- pendencies representing relatedness between task instances and correlations between infor- mation types of the tasks. However, the cross- task dependencies in prior work are not optimal as they are only designed manually according to some task heuristics. To address this issue, we propose a novel model for JointIE that aims to learn cross-task dependencies from data. In particular, we treat each task instance as a node in a dependency graph where edges between the instances are inferred through information from different layers of a pretrained language model (e.g., BERT). Furthermore, we utilize the Chow-Liu algorithm to learn a dependency tree between information types for JointIE by seeking to approximate the joint distribution of the types from data. Finally, the Chow-Liu dependency tree is used to generate cross-type patterns, serving as anchor knowledge to guide the learning of representations and dependen- cies between instances for JointIE. Experimen- tal results show that our proposed model sig- nificantly outperforms strong JointIE baselines over four datasets with different languages. 1 Introduction Entity mention recognition (EMR), event trigger detection (ETD), event argument extraction (EAE), and relation extraction (RE) are four main chal- lenging tasks in information extraction (IE), which aim to extract entities (e.g., a person), events (e.g., an attack), event arguments (e.g., a victim in an attack), and relations (e.g., work-for) mentioned ∗ Work done at Raytheon BBN Technologies (prior to joining AWS AI). in text. These IE tasks have been solved mostly in pipelined approaches (Li et al., 2013; Nguyen and Grishman, 2015; Chen et al., 2015; Lai et al., 2020; Du and Cardie, 2020; Veyseh et al., 2020; Li et al., 2020; Pouran Ben Veyseh et al., 2021; Nguyen et al., 2021b), where input to a model per- forming an IE task involves predictions from other models performing other IE tasks. As a result, er- rors in predictions by a model can be propagated to subsequent models in the pipeline to hurt overall performance. To avoid error propagation, the four IE tasks can be solved jointly (JointIE) in a single model (Lin et al., 2020; Nguyen et al., 2021a; Zhang and Ji, 2021). As such, a key challenge for JointIE models is to effectively capture dependencies between the IE …

Content truncated to 500 words